WORKING PAPERS IN IRISH STUDIES - Queen's ...

WORKING PAPERS IN IRISH STUDIES
 
 Death, Demography and the Denominator:
 Age-Adjusted Influenza-18 Mortality in Ireland
 Christopher L. Colvin (Queen’s University Belfast)
 Eoin McLaughlin (University College Cork)

 This version: 20 January 2021

Abstract: Using the Irish experience of the 1918-1919 Spanish flu pandemic (“Influenza-18”),
we demonstrate how pandemic mortality statistics can be sensitive to the demographic
composition of a country. We build a new spatially disaggregated population database for
Ireland’s 32 counties for 1911-1920 with vital statistics on births, ageing, migration and
deaths. Our principal contribution is to show why, and how, age-at-death data should be used
to construct the age-standardised statistics necessary to make meaningful comparisons of
mortality rates across time and space. We conclude that studies of the economic
consequences of pandemics must better control for demographic factors if they are to yield
useful policy-relevant insights. For example, while Northern Ireland had a higher crude death
rate during the first wave of the Covid-19 pandemic, it also has an older population; age-
adjusted mortality paints a very different picture.
Subject areas: economic history, demographic history, population economics.
Correspondence to: chris.colvin@qub.ac.uk.

Death, Demography and the Denominator:
 Age-Adjusted Influenza-18 Mortality in Ireland*

 Christopher L. Colvin† Eoin McLaughlin‡

 Original Working Paper: 26 May 2020
 Initial Journal Submission: 12 November 2020
 Manuscript Accepted: 20 January 2021

 Abstract:

 Using the Irish experience of the 1918-1919 Spanish flu pandemic (“Influenza-18”), we demonstrate
 how pandemic mortality statistics can be sensitive to the demographic composition of a country. We
 build a new spatially disaggregated population database for Ireland’s 32 counties for 1911-1920 with
 vital statistics on births, ageing, migration and deaths. Our principal contribution is to show why, and
 how, age-at-death data should be used to construct the age-standardised statistics necessary to make
 meaningful comparisons of mortality rates across time and space. We conclude that studies of the
 economic consequences of pandemics must better control for demographic factors if they are to yield
 useful policy-relevant insights. For example, while Northern Ireland had a higher crude death rate
 during the first wave of the Covid-19 pandemic, it also has an older population; age-adjusted mortality
 paints a very different picture.

 Keywords: demographic economics, pandemics, age-adjusted mortality, Spanish flu, Covid-19, Ireland.

 JEL Codes: N34, I18, Q54.

* We thank: Kyle Richmond for excellent research assistance; Robin Adams, Graham Brownlow, Matthias Blum,
Edmund Cannon, Alan de Bromhead, Áine Doran, Alan Fernihough, Hampton Gaddy, Liam Kennedy, Patricia
Marsh, Ida Milne, Cormac Ó Gráda, Florian Ramseger and John Turner for feedback on drafts of this paper;
participants of webinars organised by the Centre for Economic History at Queen’s University Belfast (June 2020) and
the new Pandemics History Research Network (September 2020) for allowing us to workshop our ideas; and the editor
(Jörg Baten) and two referees for helping us to correct errors in our initial submission to this journal.
† Queen’s University Belfast. ORCID: https://orcid.org/0000-0002-8140-0115. Email: chris.colvin@qub.ac.uk.
‡ University College Cork. ORCID: https://orcid.org/0000-0002-0307-3374. Email: eoin.mclaughlin@ucc.ie.

1. Introduction
Lessons from the Great Depression were widely employed by economists and policymakers to
understand and respond to the Great Recession (Eichengreen 2012). As the world’s last truly
global pandemic before Covid-19, the influenza pandemic of 1918-1919 (“Influenza-18”) is now
being used in much the same way. A plethora of new studies has already emerged which use this
historical pandemic to discuss the potential short- and medium-term social, economic and
political ramifications of Covid-19 (see, e.g., Barro et al. 2020; Basco et al. 2020; Benmelech and
Boberg-Fazlic et al. 2020; Frydman 2020; Carillo and Jappelli 2020; Chapelle 2020; Correia et
al. 2020; Dahl et al. 2020; Lilley et al. 2020; Lin and Meissner 2020; Velde 2020).
 But learning lessons from history is hard, and the history profession has long shied away
from doing so out of fear that practically anything can be justified by appealing to one or other
interpretation of past events (Colvin and Winfree 2019). The advantage economists had in 2008
was that we knew a lot about the Great Depression; indeed, the field of macroeconomics
ostensibly originated in that very crisis. We do not have the same home advantage when we
look to Influenza-18 to draw our lessons; economists are relatively new to this topic and run the
risk of making rookie errors which distort our findings and could lead to poorly designed policy
advice.
 We highlight one such error using the case study of Ireland, which in 1918 was still an
impoverished region of one of the world’s most advanced economies: the United Kingdom of
Great Britain and Ireland. We show that fragile population estimates combined with a failure
to adjust mortality statistics for age-at-death have distorted our understanding of this historical
pandemic. Just as currencies need to be adjusted for purchasing power when comparing the
output of different economies, mortality statistics need to undergo age standardisation when
comparing the impact of pandemics across different populations.
 Recent debates in the Proceedings of the National Academy of Sciences (Dowd et al.
2020a; Nepomuceno et al. 2020; Dowd et al. 2020b) and Science Magazine (Nordling 2020) call
for greater use of demographic methods in the study of Covid-19’s mortality burden. Meanwhile,
medical scholars are reminding their peers about the importance of measuring excess mortality
rather than cause-specific mortality when gauging the impact of Covid-19 (e.g., Beaney et al.
2020). The economics community now needs to take stock of these and related discussions. Some
models of the macroeconomic impact of pandemics already incorporate historical epidemiology

 1

(e.g., Keogh-Brown et al. 2010a,b). They should now also make adjustment for demographic
change. In this paper, we prescribe off-the-shelf solutions from the field of demography, which
can be easily employed in economic research into Influenza-18 and Covid-19 alike.

2. Why Demography Matters
Influenza-18 was a highly infectious virus with significant human and economic costs. While its
case fatality rate was relatively low, the Spanish flu, as it is still commonly known, may have
had a global death toll upward of 50 million (Jester et al. 2019; Jordan et al. 2019).1 An estimated
2.6 million deaths occurred in Europe alone – 1.1 per cent of the continent’s population (Ansart
et al. 2009). While their ultimate cause of death is generally thought to have been the influenza
A virus subtype H1N1, typical proximate causes of death during this pandemic were secondary
bacterial pneumonia infections and respiratory failure (Taubenberger 2006) – much as in the
case of SARS-CoV-2 (ECDC 2020).
 Figure 1 plots the crude mortality rate of the constituent polities of the UK, alongside
that of other countries, at an annual frequency between 1900 and 1920.2 Figure 2 reports crude
excess mortality for Ireland at a quarterly frequency across the same period. These two figures
reveal the severity of the pandemic, but they also highlight the fact that Influenza-18 took place
within a high mortality environment where outbreaks and epidemics were more commonplace
than today. Indeed, this was a time of epidemiological transition, when infectious (exogenous)
diseases remained a more significant cause of death than chronic lifestyle (endogenous) causes
(Omran 1971; Dyson 2010).
 Influenza-18 originated from an unknown source (Taubenberger et al. 2001), and
typically arrived in countries in major trading ports – carried, among others, by military
personnel. The disease diffused through populations in a process of spatial contagion along major
transportation infrastructure, typically in multiple waves (Smallman-Raynor et al. 2002). Great
Britain and Ireland saw three waves of disease, with the second being the deadliest
(Honingsbaum 2009; Milne 2018). But while all areas of the world saw excess mortality, rates

1
 There is significant disagreement in the literature on the global mortality burden of the Spanish flu, with popularly-
cited estimates of between 24.7 and 39.3 million at one end of the spectrum (Patterson and Pyle 1991), and between
50 and 100 million at the other (Johnson and Mueller 2002). An indirect estimation of all-cause pandemic-related
mortality by Spreeuwenberg et al. (2018) revises the death toll significantly downwards, to 17.4 million.
2
 Note that differences in death certification standards between jurisdictions make international comparisons
challenging (see Alderson 1981, pp. 34-40, for discussion of different registration systems across countries). This is
true also today (see, e.g., Martin 2020).

 2

differed significantly across countries (Murray et al. 2006); Influenza-18 proved particularly
deadly in developing countries (Chandra et al. 2012). There was also significant heterogeneity
in the flu’s health impact within countries, an outcome driven by local differences in demography,
density, economy, environment and policy (Patterson and Pyle 1991; Hatchett et al. 2007; Clay
et al. 2018, 2019).
 Age mattered. Indeed, worldwide the conventional wisdom is that Spanish flu was
particularly fatal to those aged between 20 and 40 years (Johnson and Mueller 2002). One
hypothesis is these young adults were immunologically naïve, with older age groups benefitting
from inbuilt immunity due to exposure to previous influenza outbreaks (Palese 2004;
Taubenberger 2006). We can also speculate that young adults, who are more likely to be in
employment, are inherently more at risk of catching the disease in societies which do not
implement “lockdown” measures (cf. Hatchett 2007). Sex also mattered; men tended to be
affected more by the disease than women (Noymer and Garenne 2000). Male flu victims
represented a population which in the absence of a pandemic would have been more at risk of
contracting high-mortality endemic diseases, like tuberculosis (Noymer 2012).
 Mortality statistics are widely used to track the severity of the Spanish flu and other
pandemics. Because of the specific demographic factors which influence the course of any disease,
academic discussions should, wherever possible, be making use of age-adjusted mortality rates
based on vital statistics collected through civil registration. They should disaggregate their
standardised mortality rates by sex, race and other categories of relevance. Direct comparisons
within and between countries are difficult to make in the absence of such statistics. Any
inferences drawn about the efficacy of public health policy, the speed of economic recovery or
the electoral consequences of lockdown measures where the underlying demographic structures
have not been accounted for cannot be robust. This is best illustrated with an example: if one
location has a higher mortality rate than another, and also has a higher share of the population
between 40-60 years of age, or a greater proportion of women, then we need to adjust for these
differences because this location’s population is inherently less exposed to Influenza-18 than the
other (i.e., it has a lower biological risk factor).3

3
 A relevant modern comparison used in Dowd et al. (2020) is that of Brazil and Nigeria. In Brazil today 21% of the
population is under the age of 15 and 10% over 65; in Nigeria 44% is under-15 and 3% is over 65. If a disease, like
Covid-19, disproportionately affects the old and leaves the young unharmed, then comparative discussions of its
impact must consider that Brazil’s population is inherently more vulnerable than Nigeria’s.

 3

Although we have long known that demography matters, the literature on Influenza-18
tends to rely on fragile population estimates, raw death counts and crude mortality rates. Most
of last year’s new crop of pandemic economics studies make no acknowledgement of, let alone
adjust for, the demographic composition of the countries and regions under study, either for
1918 or today (e.g., Lin and Meissner 2020). Even those studies which do take note of the
differential demographic impact fail to consider its economic consequences.4 Many scholars even
appear unaware that their chosen population statistics are arithmetic interpolations rather than
real population counts (e.g., Correia et al. 2020). Studies that do incorporate population change
fail to acknowledge the limitation of the underlying estimation methodology (e.g., Lilley et al.
2020). Beach et al. (2020), a new review of all aspects of the literature on the Spanish flu, largely
overlooks issues stemming from demography. But a closer reading of some of the data sources
used by researchers in this field would have highlighted the problem; contemporaneously it was
acknowledged that ‘owing to recent unusual migrations of the population and the fact that 1916
[1917, 1918, 1919] is far away from the last census year, the [population] estimates are probably
too high in some cases and too low in others’ (US CB 1918b, p. 62; 1919, p. 77; 1920, pp. 118-
119; 1921, pp. 118-119).5
 Why are economists ignoring demography? We think the answer lies with an apparent
unawareness amongst our ranks of the standard statistics, and the data required to calculate
these statistics, that demographers and health scientists use to understand mortality risks. We
think this blind spot has led to scholars making use of “convenient” data, typically derived from
the closest previous census. But this choice only amplifies the problem; we demonstrate that the
use of census data distorts the denominator in mortality statistics – especially at a time of
national crisis, such as during a war. We measure the consequence of their methodological choice
using the case of Ireland. We show that choosing the convenient denominator means scholars
fail to take full account of demographic change during World War I, a war in which large
numbers of Irish men were sent to their deaths on the battlefields of Etreux, Gallipoli, Ypres,
Hulluch, Passchendaele, the Somme and others.
 Using what demographers call the “cohort component method”, which involves ‘a
separate analysis of the changes affecting each component of the population’ (UN DESA 1956),

4
 Correia et al. (2020), for example, exploit regional variation in Spanish flu mortality in the US to look at the
pandemic’s subsequent economic impact, but do not control for regional heterogeneity in population structure.
5
 After the 1920 US census, estimated growth by state was re-estimated as ‘the populations are estimated by the
arithmetical method based on the 1910 and 1920 censuses’ (US CB 1922, p.74).

 4

we construct a new spatially disaggregated demographic dataset for the case of Ireland. Our
analysis of this dataset suggests the very youngest in society, those under the age of five, had
the highest mortality rates during the Spanish flu pandemic. Rather than focusing solely on
young adults, our attention should therefore also be drawn to young children. Development and
health economists have long pointed out that populations which survive catastrophic risks are
selected populations with fundamentally different attributes to pre-crisis populations (Deaton
2007; see review in Blum et al. 2020). This insight may also apply to our case; the elimination
of a cohort of young children may have fundamentally changed the attributes of Ireland’s
surviving population into adulthood (cf. Almond 2005; Brown and Thomas 2018).
 We think this finding has potentially significant implications for economic analyses
linking mortality rates with other datasets, in this and other settings, historical or present day.
Policymakers are currently making difficult choices about non-pharmaceutical public health
interventions, designing targeted vaccination programmes, or weighing up the costs and benefits
of financial support to firms, industries or regions. Our 100-year-old historical analogue
demonstrates that carefully constructed population data and sensible demographic adjustments
are necessary for our work to be useful to them. Health scientists are already starting to make
such adjustments in their analyses of the Covid-19 pandemic (see, e.g., Bhopal and Bhopal 2020;
Dowd et al. 2020; Kulu and Dorey 2020); economists must now follow suit.

3. The Irish Case
The impact of the Spanish flu on Ireland was first quantified by Ireland’s then-Registrar-General,
Sir William John Thompson, in an article published immediately after the pandemic’s conclusion
(Thompson 1919). Thompson limited his scope to influenza and pneumonia as the two causes of
death associated with the pandemic.6 He estimated an influenza mortality rate of 243 per 100,000
in 1918, with urban areas experiencing a rate of 370 per 100,000. He ascribed to the Spanish flu
45 of the 140 per 100,000 who died of pneumonia. He calculated this excess mortality rate by
comparing pneumonia deaths in 1918 with 1917. This brought Ireland’s total N1H1-related
mortality rate to 288 per 100,000 for 1918.

6
 Ireland’s Registrar-General adopted a more straightforward methodology than his contemporary counterparts in
London for their quantification of the mortality burden in England and Wales, who also included bronchitis, heart
disease and phthisis (BPP 1920a).

 5

The Irish experience of the Spanish flu was brought back to life again ten years ago in
three PhD dissertations completed in close succession: by Caitríona Foley (examined at
University College Dublin in 2009), Patricia Marsh (Queen’s University Belfast in 2010) and Ida
Milne (Trinity College Dublin in 2011). Foley’s dissertation was subsequently published as a
monograph, in 2011; Milne’s was published in 2018. Foley’s (2011) medical history-focused work
revolves around how “ordinary people” reacted to the pandemic. She also recounts how Ireland’s
medical professionals understood the science of infection and catalogues the various treatments
they used. An exciting feature of her work is that she puts the Spanish flu into its long-run
historical context, highlighting the fact that Ireland regularly suffered epidemics, including the
so-called Russian flu in the early 1890s (Foley 2011, chap. 2).
 Milne’s (2018) social history of the Spanish flu is a rich description of life during the
pandemic, including details on the public policy response, particularly in the eastern province of
Leinster. Her description of Ireland’s medical infrastructure highlights how the country’s local
funding model for healthcare provisioning was incapable of dealing with a national health crisis.
She describes how Irish officials had recently adopted international conventions on the collection
and classification of health statistics, making Ireland’s official statistics comparable with other
countries. However, as was the case elsewhere in the world, she notes how doctors struggled to
define cause of death; many individuals recorded as dying of tuberculosis, bronchitis, heart failure
and other maladies probably died of the Spanish flu. Official statistics therefore likely
underestimate the pandemic’s true impact, she argues. Milne estimates that 23,000 people died
in the pandemic in Ireland, from which she infers one-fifth of the island’s population probably
contracted the disease (Milne 2018, chap. 3). Using weekly data available only for the Dublin
registry district, she calculates that the death rate for the city’s poorest social classes was
probably almost double that of the richest (Milne 2018, chap. 3).7
 Also very relevant to our research is Marsh’s (2010) unpublished dissertation on the
Spanish flu in the northern province of Ulster. Of the three PhDs it includes the most extensive
quantitative, demographic analysis. Marsh uncovers the precise timing of each of the three waves
of the pandemic using a combination of official sources and local newspapers. Subsequently, she
exploits the same official government sources we use in this current paper, and adopting the
more sophisticated methodology of the then-Registrar General of England and Wales she re-

7
 Elsewhere, Milne (2019) describes how she collected oral histories of the pandemic, which she argues add an
understanding which ‘cannot be discovered or quantified by the Registrar-General’s statistics’ (p. 16).

 6

calculates excess mortality statistics for Ireland using a broader list of causes of death than
Thompson. She is therefore able to improve comparisons with the other two health statistics
jurisdictions of the UK (England and Wales, and Scotland). Marsh estimates there may have
been up to 14,000 additional deaths than previously ascribed to the flu, taking the total pandemic
death toll to 34,000, and yielding a crude mortality rate of 782 per 100,000 population.8
 The only work to make use of Irish influenza statistics in an econometric analysis is de
Bromhead et al. (2020). The article is an analysis of the 1918 general election, which took place
on 14 December – at the end of the second wave of the pandemic – and gave the previously-
obscure Sinn Féin party the majority of Ireland’s 105 Westminster MPs. Using the same official
sources we use, de Bromhead et al. calculate crude Influenza-18 mortality rates at the Poor Law
Union level, the lowest administrative division at which health statistics were reported. They
then use GIS software to allocate these to electoral constituencies – not an easy task since the
two sets of boundaries do not line up. They find higher Spanish flu mortality is associated with
a lower turnout on election day, but they argue this did not affect the overall electoral outcome.
 To calculate the mortality rate, a demographic researcher requires a denominator: the
relevant population, the average population exposed to risk of death during the defined time
period. Up-to-date population estimates are necessary for a variety of administrative indicators,
such as vital statistics and disease incidence (Long 1993). For their denominators, Thompson,
Milne, Marsh and de Bromhead et al. all use the population taken from a census of Ireland
conducted on 24 April 1911 (BPP 1913a). This choice is understandable as the most
comprehensive sources for demographic data are censuses, and 1911 was the most recent census
year.9
 But populations can change, sometimes quite suddenly. Relying solely on the 1911 census
means Thompson and the others do not take account of changes in population due to births,
ageing, migration and deaths since 1911. Notably, the 1911 census fails to take account of the
falling birth rates, the ageing population and changes in the composition of deaths due to
warfare. The outdated denominator is compounded by a failure to age standardise Influenza-18
mortality statistics.

8
 Elsewhere, Marsh (2019) further discusses some of her PhD’s findings on gender and mortality in the case of Ulster.
She argues Ulster’s higher female fatalities may have been driven by the province’s higher female-to-male ratio.
9
 This census was based on a house-to-house collection of data and was considered by census officials to be their best-
yet.

 7

Irish scholars of the Spanish flu are not alone in their choice to use the closest census.
Around the world the most recent census year to the 1918-1919 pandemic were typically around
1910 (ex ante/pre-pandemic) or 1920 (ex post/post-pandemic), and so this same “error” is being
repeated across the entirety of this literature. Age standardisation does not feature in any of last
year’s new Influenza-18 studies.

4. Age Adjustment and Standardisation
To quantify the demographic impact of the Spanish flu, we need to estimate a denominator for
our mortality statistics; relying on the closest pre- or post-pandemic census years will distort the
true effects of the pandemic. At the most basic level, the population has changed in terms of
births, ageing, migration and deaths in the intervening period. This was a very turbulent era,
with the outbreak of a global war in August 1914 which lasted until November 1918, and a
rebellion in Dublin City in April 1916; the 1911 census cannot take these into account. The
pandemic was closely followed by a revolution, a guerrilla war, political partition of the island,
large population resettlements of Protestants to Northern Ireland and a civil war in the newly-
formed Irish Free State; the next censuses, conducted simultaneously in both Irish jurisdictions
in 1926, is also not an appropriate choice for our denominator.
 We need to make our own postcensal estimates of the population in 1918 and 1919 to
calculate influenza-related mortality rates in 1918 and 1919:

 (1)
 = × 100,000
 
where is the mortality rate per 100,000 population in year t; is the number of deaths in
year t; and the base population in year t, typically measured mid-year. Because the nature of
the disease means men and women were impacted differently, we need to calculate this separately
by sex. After sex, age is the single most crucial variable when studying mortality (McGehee
2004); we need an accurate picture of the age structure of the population. Only once we have
population by sex at each age, are we able to calculate age-specific mortality rates:

 (2)
 = × 100,000
 
where is the mortality rate per 100,000 population for sex s at a specific age or age-range
a in year t; is the number of deaths for sex s at that age or age-range a in year t; and 
the base population for sex s in that age or age-range a in year t, typically measured mid-year.

 8

Alternatively, the crude death rate can be calculated as a weighted average of the death-by-age
( ) and the population share of the relevant demographic ( ):

 ∑ (3)
 = × 100,000
 
 Demographers have long observed that comparisons of mortality using crude death rates
alone can be misleading as the demographic (sex, age, race) composition of a population will
affect the level of the observed death rate (see, e.g., Linder and Grove 1947, p. 60). Although
the crude mortality rate is a weighted average mortality based on the age composition of a given
population, we need to take account of age differences in order to make comparisons across
populations with different age distributions. We can do this by imposing group weightings from
a “standard” population. Doing this creates a hypothetical death rate which assumes the
demographic composition of the population under study equals that of the standard population.
 We calculate several estimates of a standard population. In order to make contemporary
comparisons, we calculate a weighted average standard population for 1911 using data on OECD
age structures that were reported in the 1911 Census of England and Wales (BPP 1917a, p.
63).10 We also compare these with four more modern standard populations: 1940 (US Standard
Million), 1960 (World Standard Million), 2000 (WHO Standard Population), and 2011 (EU27
Standard) (NCI 2012; Ahmad et al. 2001).11 We report these standard population weights in
Table 1 for the relevant reporting age bands, alongside our own estimates for modern Ireland
for comparison.12 The earlier age structures are more relevant for Influenza-18, when populations
were younger on average; the latter standard populations are most relevant for Covid-19, where
developed countries have populations that are older than developing countries on average.
 Here we adopt a direct age standardisation approach using:

 ∑ (4)
 = × 100,000
 
where refers to the mortality rate of the specific population under study; is the standard
population at each age; and P is the total standard population. Effectively this means we
calculate a weighted average age-specific mortality rate.

10
 For OECD countries, we take: Australia, Austria, Belgium, Canada, Denmark, England and Wales, France,
Germany, Ireland, Italy, the Netherlands, Norway, Portugal, Scotland, South Africa, Sweden, United States.
11
 A selection of World Standard Population data are available from the WHO website
(https://apps.who.int/healthinfo/statistics/mortality/whodpms/definitions/pop.htm).
12
 Spatially disaggregated (county-level) population weights are reported in Colvin et al. (2021).

 9

In the existing Spanish flu literature there is a discrepancy over the term “age
adjustment”. For example, Taubenberger (2006, p. 94) refers to ‘age-adjusted mortality data’,
but the data that he refers to are actually age-specific mortality rates (death rates by different
age bands); this is the same approach adopted by Milne (2018) for the case of Ireland. Age-
specific mortality is useful for looking at a national picture. But demographic composition will
distort spatial and temporal comparisons, so age standardisation is necessary; age adjustment
needs to take account both the age-specific mortality and make an adjustment for differences in
population structures by applying relevant weights to age-specific mortality estimates.

5. Cohort Component Estimation
Ireland’s system of registration of births, marriages and deaths commenced in 1864.13 Annual
reports collated registration statistics from 130 reporting districts (Poor Law Unions) throughout
the country and aggregated these to 32 counties, which are then further collapsed into the
island’s four historic provinces (Leinster, Munster, Ulster and Connacht). We rely on these data
as our sources for (the number of deaths in year t). The data were collated by age of death,
sex, and cause of death. We digitised all Registrar General Reports between 1911 and 1920 to
estimate our postcensal populations to use as the denominator in our mortality statistics (BPP
1912b, 1913c, 1914b, 1915b, 1916c, 1917c, 1918b, 1919b, 1920c, 1921b). The base population for
our age structure estimates comes from the 1911 census (BPP 1913a), which was digitised by
Clarkson et al. (1997).
 To estimate the population in 1918 and 1919, we then use the cohort component method
from demography (e.g. Bryan 2004; Watcheter 2014), which takes account of vital statistics in
the intervening years (1911-1918). Essentially, we update the census with data on population
flows. We use the following equation:

 = ( ) + ( − + − ) (5)

where is the population for sex s at age or age-range a in county c at time t; and B, D, I
and E are within-year births, deaths, immigration and emigration, respectively.
 The simple and cohort component methods of estimating population change have been
considered the “gold standard” in population projection since at least World War II (UN DESA

13
 Walsh (1970) compares the registration of births with census reported births and notes that registration improves
significantly by the 1910s.

 10

1956), and they are still widely used by statistical agencies to estimate population (e.g., US CB
2014; ONS 2020).14 Population projections have their limitations and are not without
controversy; demographers have missed post-war baby booms and subsequent baby lulls (De
Gans 2002; Watcheter 2014). However, we do not suffer from this problem, as we use the method
in a short, contained, timeframe to estimate historical populations and do not make assumptions
about past trends.
 We use annual records to help estimate postcensal populations between two decadal
censuses at a spatially disaggregated level, along the lines described by Long (1996). National-
level population estimates were estimated annually by contemporary government statisticians
using a simple component method and were reported in the Registrar General Reports. Their
methodology is straightforward: ‘by adding the births registered in each year to the estimated
population for the previous year, deducting the deaths and the number of emigrants’ (BPP
1921b, p. 40). Essentially, we follow this same procedure at a district level for Ireland’s 32
counties, while also adjusting for ageing and migration.
 We use direct measurement of births, deaths, and emigration. Contemporary
immigration statistics are, unfortunately, harder to procure. Instead, we use information on
annual immigration figures derived from Social Welfare (1955, p. 326).15 We focus here on
international rather than internal migration.16 Deaths are recorded by age range in Ireland’s
vital statistics, and so we can account for deaths within defined five-year age bins. Emigration
age profiles are available from contemporary records (BPP 1912a, 1913b, 1914a, 1915a, 1916a,
1917b, 1918a, 1919a, 1920b). Immigration was considerably smaller than emigration in this
period, a consequence of global warfare, so effectively this is a net migration story.

14
 The same method is used for sub-national population projections by the statistical agencies of the Republic of
Ireland (https://www.cso.ie/en/methods/surveybackgroundnotes/populationandlabourforceprojections) and
Northern Ireland (https://www.nisra.gov.uk/publications/2018-based-population-projections-areas-within-northern-
ireland).
15
 These are mostly return migrants from emigrant destination (see Fernihough and Ó Gráda 2019). Ireland
traditionally experienced net emigration, the main reason for sustained population decline over time. Pre-war
immigration was 20% of the emigration total. We assume immigration follows a similar spatial pattern as emigration.
16
 Census figures provide details of location of birth of the population. We can estimate internal migration by
comparing the 1901 and 1911 censuses (BPP 1901; 1913a). In 1901, 11.38% of the population of counties was born
outwith the county; in 1911, this figure rose to 13.20%. This equates to a growth of 0.18% per annum. The outliers
in this were the major urban centres (Belfast and Dublin), with the lowest rate of internal migration seen in the west
(Kerry, with 3.85% in 1901 and 3.96% in 1911). Internal migration tended to be to the nearest urban centre, so
province- rather than county-level statistics essentially already incorporate internal migration.

 11

To estimate and annually update the age structure of the population, we move a share
of each age bin forward one year, assuming a survival rate for the last year in the cohort. We
explain the procedure for the age bin covering 25-29 year-olds as an example of our methodology.
First, we account for all deaths and migration flows. Then, we bring a share of the population
forward to the next bin; the 29-year-olds are moved to the 30-34 bin and are replaced in the 25-
29 bin by 24-year-olds from the 20-24 bin. We adjust for the 24- and 29-year-old survival rates
using weights calculated from the same age band in the 1926 census, the next census of Ireland
following the 1911 census.17
 We collapse all ages over 65 into one single age band as there are perceived discrepancies
with age statements in the 1911 census, our base year. Ireland’s census commissioners believed
the 1911 was more accurate than the previous census in terms of age statements as there was
less heaping (reporting of rounded ages at 0 or 5) in the final report.18 The commissioners believed
older people had in the past under-stated their age, but because of the recent introduction of
the 1908 Old Age Pension for over-70s, they had for the first time ‘ascertained their correct age’
(BPP 1913a, p. 25). Ireland’s long history of youth emigration had left the country with an older
population (Akenson 1993; Fitzpatrick 1980, 1984). It is likely that over-65s were therefore
already a disproportionate share of the population, for demographic rather than nefarious
reasons. However, Budd and Guinnane (1991) fear there was a deliberate overstatement of age
to qualify for the pension. Collapsing the age bins over 65 years of age is our way to circumvent
the possibility of overstatement of ages, while not trying to manipulate the underlying census.
 Military enlistment during World War I is effectively treated as emigration in the
Registrar General’s population estimates as it was a sizeable population movement to the
battlefields of Europe. Subsequent demobilisation was treated as return migration (immigration).
For example, in the population estimates for 1914, the male population decreases by 49,881, and
in 1919 it increases by 63,000 (BPP 1921b). We follow a similar procedure. To adjust for military
enlistment during World War I, we estimate the total enlistment in Ireland (134,202) from
contemporary military sources (BPP 1921a, p. 9; WO 1922, p. 363).19

17
 This census was conducted separately for the by then jurisdictionally partitioned island (Government of Northern
Ireland 1929; Roinn Tionscail agus Tráchtála 1928, 1929).
18
 See Blum et al. (2017) for a discussion of age heaping in Irish census data.
19
 Our enlistment figure is considerably lower than Bowman’s (2014) recent estimates, who believes 210,000 Irish
served in the British Army. This discrepancy is because the Army figures of enlistment for Ireland explicitly exclude
‘Irishmen enlisted in Great Britain who came over for the purpose’ (BPP 1921a, p. 6).

 12

We detect the county composition of Irish enlistment from a further parliamentary source
which covers 97 per cent of the total Irish recruits (BPP 1916b). The age of military service was
between 19 and 41, but government statisticians noted ‘the male population of Ireland is
composed chiefly of young men up to 18 years of age and of men over 50, as a large proportion
of the remainder emigrates to the United States and Colonies’ (BPP 1921a, p. 9). Estimated
military mortality is derived from War Office statistics, which imply a mortality rate of 14 per
cent for all military personnel. Our war-related mortality estimate is lower than Bowman’s
(2014) estimated 20 per cent mortality.20 We have chosen the more conservative death rate of
14 per cent to keep Irish mortality in line with UK-wide mortality figures.21
 For cause-specific death rates in the pandemic, we report data on deaths attributed to
influenza in addition to those attributed to all forms of pneumonia. While the inclusion of
pneumonia with influenza deaths was common internationally, such as in the US (see US CB
1920, pp. 29-32; and used, e.g., in Brainerd and Siegler 2003), there are limitations of the data
due to mistakenly attributing deaths to other causes, notably, to other respiratory illnesses (cf.
Milne 2018, chap. 3). However, a no-pandemic counterfactual would also have seen many deaths
due to influenza and pneumonia; despite being a great power, the UK remained a high-disease
environment with a poor public health infrastructure, and so counting all who died of these two
maladies would over-estimate the impact of Influenza-18. Equally, year-on-year fluctuations in
the background mortality of certain age groups driven by other non-pandemic factors may result
in an under-estimate of the pandemic’s true impact (Andreasen and Simonsen 2011).
 To overcome these problems, we also report three excess mortality statistics: (1) deaths
from all causes relative to what would “normally” have been expected across a given year; (2)
deaths from influenza and pneumonia relative to normal; and (3) deaths from all causes for those
aged under 45, the most severely-affected demographic group, relative to normal for that
population group. Our estimates are based on a comparison of the mean influenza and pneumonia
death rates for 1910-1914 with those in 1918 and 1919, calculated by sex and age. The selection
of pre-war mean death rates is an attempt to not distort estimates with wartime spillovers.
Moreover, 1915 saw a higher than usual number of deaths from tuberculosis, so including this
year would distort the comparison. We report age-adjusted excess mortality using the estimated

20
 Bowman’s (2014) estimated mortality is 5.7% of all reported UK casualties, although his enlistment estimates
suggest Irish comprised 4% of the British Army, implying Irish soldiers were over-represented in mortality figures.
21
 The average difference between the sum of our county population estimates and those estimated by the Registrar
General is -0.26%. We attribute this difference to our treatment of troop movements.

 13

populations in 1918 and 1919 to account for changes in the composition of the population in the
intervening period.
 Our third measure of excess mortality follows an approach proposed by Andreasen and
Simonsen (2011). The under-45s are well-documented across Influenza-18 studies to have suffered
a disproportionate mortality burden. Consequently, Andreasen and Simonsen argue that
including the over-45s in excess mortality calculations runs the risk of under-estimating the total
burden of the pandemic because the background mortality fluctuations in older age groups may
outweigh any impact of the pandemic. Excluding the over-45s means our excess mortality
statistics are focused only on the population most at risk of pandemic-related death.22
 Our full underlying cohort component population estimates for Ireland are available in
a separate data paper (Colvin et al. 2021). Note that there exist various alternative
methodologies for estimating population that are used to make population projections. One such
alternative is linear interpolation. This method was used by US census officials to ensure timely
estimates could be computed; ‘the method of arithmetical progression was adopted for computing
the estimates of population […] based on the assumption that the increase each year since the
enumeration is equal to the annual increase from 1900 to 1910’ (US CB 1918a, p. 5).23 Adopting
the interpolation methodology, given the data we now have available, would be rather crude and
would hide variation within Ireland.24 Especially important in our case is that linear interpolation
would yield misleading age compositions as it assumes trends from the previous decade are
constant.25

6. Findings
We report our full county-level mortality statistics in the Appendix (Tables A1 and A2). Our
starting point is a comparison of crude death rates across time, using Table 2 and Figure 1.
Ireland’s crude death rate increased from 16.52 per 1,000 in 1911, to a peak of 18.15 in 1918.

22
 A similar argument can be made for Covid-19 mortality statistics, which we think should be consistently reported
for the over-80s alongside the cruder measures typically used by journalists.
23
 Indeed, it is these linear interpretations of population that are used in recent economic studies of the Spanish flu
(e.g., Correira et al. 2020).
24
 At a national level the linear interpolation using the 1901-1911 population growth is close 0.15% per annum, versus
0.18% for 1911-1918 using the component method.
25
 Take two examples: Leitrim’s growth rate between 1901-1911 was -0.86% per annum, while the population growth
was effectively zero between 1901-1918 owing to a reduction in emigration. Galway also experienced negative growth
between 1901 and 1911 (-0.55%), but experienced growth between 1911 and 1918 (0.15%) owing to falling emigration.

 14

Here we see a marked difference between Influenza-18 and Covid-19: the prevailing levels of
mortality are higher as the Spanish flu pandemic occurs in an era of epidemiologic transition.26
 Another discernible aspect of Figure 1 is the differences in the heights of the mortality
spikes in 1918. One of the key distinctions between Ireland and these comparator countries is its
unusual history of population decline from the 1850s onwards. The main driver of this decline
was emigration; Ireland had the highest emigration rates in the world (Hatton and Williamson
1994, Tab. 1.1). These emigrants were young, and the residual population was therefore older.
The inverse is true for immigrant-receiving countries, such as Australia or the US.
 Figure 2 makes use of quarterly returns of deaths in Ireland to estimate quarterly excess
mortality by comparing quarters with average mortality in the same quarter of the previous four
years. Here the striking aspect is the high excess mortality in the first quarter of 1900 almost
rivals that of 1918. This spike is apparently the consequence of an influenza epidemic, where
4,677 deaths were attributed to influenza (199 per cent higher than the average of the previous
four years) and 3,824 attributed to pneumonia (12 per cent higher than the average of the
previous four years). The crude influenza mortality rate for 1900 was 106 per 100,000; age-
adjusted, using the 1911 standard population in Table 1, the mortality rate was 99 per 100,000.27
 Within Ireland, there are significant demographic differences in terms of migration and
age structure. These are illustrated in the various population pyramids reported in Figure 3.
Notably, urban centres in the east of the island grew, drawing internal migrants, while the
countryside was depleted of youth. Comparing countries by age distribution – as done by
contemporary census officials in BPP (1917a) – Ireland has one of the older populations; the
share of over-55s was 16 per cent in Ireland versus an average of 12 per cent across other countries
listed.28
 The population of Ireland decreased by 1.3 per cent in the seven years from the 1911
census (see Table 2). This fall in population was driven by a number of factors: a significant fall
in birth rates, increasing mortality driven by an outbreak of pandemic tuberculosis in 1915,

26
 Comparison of crude death rates shows regions today that have undergone the second demographic transition are
50-60% of those in the 1910s (Lesthaeghe 2014). But such comparisons are somewhat misleading given changes to
countries’ demographic composition; in the US, for example, age-adjusted death rates are considerably lower than
crude death rates, and there are age-adjusted disparities between ethnic groups (see, e.g., Kochanek et al. 2019).
27
 Noteworthy here is that 64 per cent of the reported influenza deaths in 1900 were over the age of 55. However, the
1900 influenza pandemic was confined to one quarter, during the winter flu season, whereas Influenza-18 had excess
mortality over multiple quarters.
28
 The over-55 share of the population was 10% in the US and 12% in England and Wales (BPP 1917a).

 15

falling migration during the War, and military enlistment. If we ignore these demographic
changes and use the 1911 population as our denominator for 1918 – the methodological choice
made by Thompson (1919) and the others – then this implicitly, and incorrectly, assumes no
change in the demographic composition of the population.
 The changes in the composition are best illustrated through a visualisation of population
pyramids. Firstly, as an aggregated national picture, Figure 3 illustrates the compositional
changes wrought by this declining population. There were noticeable falls in all age cohorts
except the 5-9 and 55-64 cohorts (Panel A and Panel B). There is a distinct difference between
male and female population change, as males made up the bulk of military recruits.
Disaggregating to Ireland’s four historic provinces, we begin to see distinct regional patterns
distinguishing the east and west of the island (Panel C). Further focusing in on the population
in urban versus rural districts illustrates a disproportionate impact of military enlistment on
cities (Panel D).
 The raw data on influenza and pneumonia deaths in 1918 and 1919 are reported in Table
3. If we only look at the age distribution of deaths, we can confirm the “W-shaped” impact of
the pandemic, where those aged 20-40 were disproportionately affected by Influenza-18:
approximately 30 per cent of deaths hailed from people in this age group. However, this does
not make allowance for the demographic weighting of these groups. Table 4 calculates cause-
specific death rates using both the 1911 census weights (Panel A) and our 1918/1919 estimated
weights (Panel B). Across the board, the 1911 weights understate the impact of the pandemic.
Adjusting for demographic changes, the Influenza-18 story is more nuanced, and, if anything, it
aligns more with conventional mortality statistics, where the youngest and oldest display highest
levels of mortality. This pattern is more clearly evident by comparing Panel A with Panel B of
Figure 4.
 Table 5 reports distinct regional patterns in terms of age-specific and age-standardised
mortality rates.29 Leinster and Ulster, located in the east of the island, show considerably higher
age-standardised mortality than in Munster and Connacht, and are also high compared to the
national average (see Table 4, Panel B). Focus on the main urban centres also shows much
higher mortality across age groups, especially in the younger cohorts (under-5s), and age-

29
 In our maps we report Dublin City as part of Dublin County, and Belfast as part of Antrim. Deaths attributed to
these cities may otherwise over-report the true impact of the pandemic because the main hospital infrastructure for
these cities’ immediate hinterlands was located there.

 16

standardised mortality rates in urban centres are more than double that of rural populations.
This highlights the consequence of urban density for the spread of the disease. Disaggregated
county maps of age-adjusted influenza-related mortality are presented in Figure 5, highlighting
the considerable variation across the island.
 Age-specific excess mortality, reported in Figure 6, shows a similar pattern to age-specific
influenza and pneumonia – with the exception that the oldest cohorts recorded fewer deaths in
1918 than they had recorded (on average) between 1911 and 1914, thus giving them negative
excess mortality (see Figure 7 for a map of age-adjusted excess mortality). Excess mortality
more explicitly highlights the consequence of using the 1911 census rather than our estimated
1918/1919 denominators. The 1911 denominator records much lower than expected excess
mortality in the oldest cohorts than the 1918 denominator.
 In contrast to the cause-specific mortality figures shown in Panel B of Figure 4, the
excess mortality figures of the prime working aged (25-34) is highest in Panel B of Figure 6. This
figure, in turn, reflects the fact that the pandemic operated in a competitive disease environment.
Indeed, comparing Panel A with Panel B of Figure 6 suggests that whether the pandemic killed
more children than usual is in some sense moot as there were many competing potential causes
of childhood mortality in this era; similar numbers would probably have died in a no-pandemic
counterfactual, but of other causes. Age-adjusted all-cause mortality shows a strong positive
correlation with age-adjusted influenza and pneumonia mortality (Figure 8).
 Age of death-specific mortality rates suggest a possible mechanism of infection may have
been family units (cf. Dowd et al. 2020a); across provinces, apart from Connacht, the ratio of
the mortality rates of under-5s and 25-34s are of a similar magnitude.30 This trend is most
pronounced for males as the female under-5 cohort has significantly higher mortality than the
25-34 cohort.31 This pattern suggests a possible confounder related to the economic activity of
males, an idea raised independently by Marsh (2010) and Milne (2018). During World War I,
part of Ireland’s labour force was deemed essential for both the war effort and for domestic
morale. Reserved occupations in rural Ireland included farmers; the pastoral nature of the
agricultural economy meant little social interaction. Reserved occupations in urban centres
related to transport and factory work (railway and transport workers, food processing,

30
 Note that intergenerational family units have also been found to associated with higher fatality rates for the Covid-
19 pandemic (Aparicio Fenoll and Grossbard 2020).
31
 This is disproportionately driven by infant mortality (under the age of one).

 17

shipbuilding and repairs), where what we would now call “social distancing” was near-
impossible.32
 The major puzzle is why those over 45 had lower than expected mortality (i.e., lower
excess mortality compared to the 20-40 cohort). One explanation in the literature is they enjoyed
some immunity thanks to previous exposure to influenza strains, such as the particularly virulent
outbreak of 1900 (see Taubenberger 2006). These earlier influenza epidemics may also have
selected the population in this age group, leaving only healthier individuals behind; the influenza
death rate amongst the under-10s in 1900 was 64 per 100,000, compared with 25 and 80 per
100,000 for the 20-34 and 35-54-year-old cohorts. An alternative explanation is these older
cohorts were less economically active and so less likely to catch influenza, or less likely to work
while recovering from influenza (cf. Milne 2018).33
 We have uncovered marked differences in demographic composition between 1911 and
1918. The most prominent divers of this change were war related. Most immediate was army
enlistment from urban centres, which reduced the male population in the military-active age
groups (20-40), the exact population identified in the epidemiology literature as being
particularly susceptible to 1918’s N1H1 strain. Indirectly this led to a drop in birth rates during
the war years. We also show rural centres were older owing to emigration trends – the specific
population group traditionally identified as being less affected by the Spanish flu.

7. Discussion
As with the Great Recession referencing the Great Depression for influence in developing policy
responses, the Great Virus of today sees continued reference to historical pandemics. The major
analogue for Covid-19 is Influenza-18. The Spanish flu is now widely studied again and referred
to both in academic and popular writing, including our own (Colvin and McLaughlin 2020a;
Colvin 2020). The historical pandemic has notably been repeatedly referenced, albeit always
incorrectly, by former US President Donald Trump (Rupar 2020; Mathers 2020). However, to
fully extrapolate relevant policy lessons, we must first develop a more nuanced understanding of
the population-at-risk in 1918. This means employing off-the-shelf methods from the demography

32
 Another property of the male population is that men deemed physically unfit for military service are among those
left behind. It is feasible this selected population was more at risk of dying in the pandemic.
33
 The heightened disease exposure of the working population is consistent with evidence from the US on modern
seasonal influenza virus transmission: Markowitz (2019) find that a one percentage point increase in the employment
rate increases the number of influenza-related outpatient health care visits by 19%.

 18

literature to take account of the changing demographic composition of populations during the
1910s, a turbulent decade across the globe.
 The existing literature on the 1918-1919 pandemic stresses the importance of age and
sex, with males aged 20-40 constituting the population particularly at risk of succumbing to the
virus. But surprisingly, the new economic studies of the pandemic mostly fail to address this
fundamental issue of demography. They do not carry out age standardisation, without which
mortality statistics across different localities are very difficult to compare. Economists must pay
more attention to the source of their historical data, and take better note of the statistical
methodologies developed by those in other disciplines to interpret these data. With the right
data, age adjustment is a computationally straightforward procedure, and relevant standard
populations are readily available.34
 Additionally, the new Influenza-18 economics studies use unrepresentative population
estimates in their mortality rate denominators. These have the potential to bias their results.
The pandemic occurs late in the census cycle and starts during a world war which resulted in a
global population upheaval. Using ex ante population estimates in mortality statistics implicitly
assumes the prevailing decade had few demographic implications; adopting ex post population
estimates overstates the burden of the pandemic as the dead are no longer counted in the
population. Wherever possible, scholars should instead spend time to estimate their own
denominators for their 1918 and 1919 mortality calculations, pieced together from readily
available vital statistics published by national statistical agencies.
 Quantitative demographic histories of Ireland end with World War I (see, esp., Guinnane
1997) and therefore do not address demographic change across the specific period necessary to
analyse the 1918-1919 pandemic. We must, therefore, carry out this analysis ourselves. Our new
population estimates are reported in a separate paper (Colvin et al. 2021). Ireland’s experience
of Influenza-18 has recently seen important contributions from social and medical historians
(most notably: Marsh 2010 and Milne 2018). We complement their analysis by updating their
figures with a more robust methodology which takes age-at-death into account. While we provide

34
 Our contribution somewhat relates to Aburto et al. (2021), who find that off-the-shelf demographic methods also
help to improve our understanding of life expectancy during a pandemic.

 19